Conditiondark/astra-cadfem-failures

Astra CAD+FEM failures v0.1

51 CADFEM-Bench v1.9 tasks on which GPT-6 Astra failed in every counted attempt. Each case keeps the task exactly as Astra received it, Astra's own code and outputs, and the evidence for the verdict. Built from saved runs by scripts/build_astra_cadfem_failures.py (no model calls).

Protocol

GPT-6 Astra (gpt-6-astra), medium reasoning, one code-generation call per task, no tools and no feedback (OpenAI Batch API). The returned main.py was run in a local sandbox (macOS seatbelt, no network, 600 s, 6 GB, CPython with CadQuery/OCP, gmsh, numpy/scipy, meshio, scikit-fem, CalculiX 2.23) and scored by the CADFEM-Bench evaluator.

A case is admitted only if Astra failed on every counted attempt. Task defects (wrong prompt data, reference errors, missing inputs) and infrastructure stops (gateway errors, our resource limits) are not counted, and a task with any recorded task defect is left out entirely. Of 86 CAD+FEM tasks attempted: 51 failed_every_attempt, 8 failed_some_attempts, 6 excluded_task_defect, 20 passed, 1 excluded_only (attempted.jsonl). Three defects in our own tasks found while building this set are listed in dataset-defects-found.json.

Evaluation scheme

1. Replay. Astra's main.py runs once in the sandbox above, with no retries or feedback. The results.json it writes is its answer (astra/<id>/execution.json, log-tail.txt).
2. Score (CADFEM-Bench evaluator/score.py; evaluator/<id>/score.json). Every requested key must be a finite number within its tolerance of the reference in evaluator/references.jsonl: relative 5 % for buckling factors, SSLV110/155 K_I, the cycloidal-drive displacements and the Wyk limit load; 10 % for AgentFEM, SSLV311 and the valve temperature drops; 5 % with a 40 K offset for the valve wall temperatures. No results.json, a missing key or a non-finite value is a failure (no answer). Every case here fails this numerical check, so the benchmark's artefact, geometry and workflow gates do not change any verdict.
3. Adjudicate, so that each failure is Astra's and not ours.
   • Buckling (scripts/adjudicate.py): within 5 % is a pass. A wrong factor is wrong_trap_confirmed when Astra's own final CalculiX deck, rerun unchanged except for loads x 0.1, gives the reference within 2 %: its geometry, mesh, supports and loads were right and only the reported mode was wrong. no_result_unparsed_skipped_mode: its script failed reading its own CalculiX table, which held only skipped higher modes (the same rerun gives the reference). no_result_before_buckling_solve: it stopped on its own error before the solve. no_result: other errors in its own code.
   • Other cases: the cause is read from Astra's log and code and classed M1-M8 (explanation). Any sign that the task was at fault is checked independently (the supplied valve BREP meshes with standard gmsh settings; an independent solid modal solve showed the MORPHWING reference was a shell model). Faults on our side make the task a defect, and it is excluded.
4. Admit only tasks on which every counted attempt failed (above).

To test another model on these cases, give it only tasks/<id>/, replay with CADFEM-Bench evaluator/sandbox_runner.py, score with evaluator/score.py and run scripts/adjudicate.py on buckling runs. The set was selected on Astra's failures, so it is adversarial to Astra: report another model's score on it as such, not as a fair comparison with Astra (use the full benchmark for that).

Composition

topic	cases
stability	43
structural static	3
fracture	3
plasticity	1
thermal	1

mechanism	cases
M1	30
M2	20
M8	1

M1 = silent wrong value (trusted the solver output without a sanity check); M2 = no answer (own script, parser or solver-handling error); M8 = refused to report after its own verification failed.

answer	cases
wrong value	30
no answer	21

Buckle-trap cases (43)

Linear buckling of bars, sections, plates, stepped and tapered members with a true first factor below 1. CalculiX 2.23 *BUCKLE skips factors below about 1 (issue 74), and Astra trusts the first factor it returns. Verdicts: 30 wrong_trap_confirmed, 8 no_result_unparsed_skipped_mode, 2 no_result_before_buckling_solve, 3 no_result. A wrong answer is confirmed when Astra's own final deck, rerun unchanged except for loads x 0.1, gives the reference within 2 % (evaluator/<id>/load-scaled-rerun.json), so its mesh, supports and loads were right. Reported factors were 1.6x to 9.0x the reference (median 5.8x). References are FEniCSx P2 + SLEPc on three meshes (evaluator-computed, not published; evaluator/<id>/reference-evidence). The 43 cases come from 9 parametric families, not 43 independent problems.

Other cases

case	topic	mechanism	answer	what happened
A-AGENTFEM-fresh_test_000	structural static	M2	no answer	own CAD sanity check used gmsh's loose OCC bounding box (z from -1.39, y from -64.8) with a 1e-4 mm tolerance and aborted; the CAD is exactly z 0-189
A-CYC-Disk-PETGCF	structural static	M2	no answer	own curvilinear second-order gmsh mesh had inverted elements (CalculiX: nonpositive jacobian in 9 elements); the script stopped on its first mesh leve
A-CYC-Ring4x6-PETGCF	structural static	M2	no answer	own refinement broke: level 2 jumped 36x (0.0467 vs 0.00128 mm) unchecked, level 3 deck rejected by CalculiX (*ERROR reading *CLOAD, node 1997); scrip
A-FCVM-PL-Wyk_Rohrstoss_ring	plasticity	M8	no answer	its own equilibrium check of the plastic solution failed (relative residual 157); refused to report a load factor
A-VALVE-transient	thermal	M2	no answer	inserted probe points into the BREP before meshing; gmsh PLC error. The supplied valve.brep meshes with standard settings (45k nodes, 1.8 s)
B-CA-SSLV110	fracture	M2	no answer	own CalculiX .frd parser found no displacement field; no answer
B-CA-SSLV155	fracture	M2	no answer	converted its gmsh mesh with meshio, which has no wedge15 type; no answer
B-CA-SSLV311	fracture	M2	no answer	own crack-tip SIF extraction raised a numpy TypeError; no answer

Layout

• cases.jsonl: public index (no reference values). tasks/<id>/: inputs as given to Astra (task.json holds the prompt).
• astra/<id>/: main.py, generation.json (usage, cost), results.json if any, log-tail.txt, execution.json, deck.json + deck-inlined.inp.gz (the CalculiX deck behind the answer, includes inlined, and its factor table).
• evaluator/: references.jsonl, cases-full.jsonl (references, Astra values, errors, all attempts, sources) and per case score.json, adjudication.json, load-scaled-rerun.json, reference-evidence/.
• attempted.jsonl, excluded.jsonl, summary.json, dataset-defects-found.json, scripts/, SHA256SUMS.

Limits

One sample per task at medium reasoning without tools; some misses may not reproduce on resampling. Failures concentrate in buckling (43 of 51), where one solver behaviour drives them; 20 cases are no-answer failures in Astra's own code, two of them parser failures on CalculiX tables whose first factor was correct. 8 cases have publisher references; the rest are evaluator-computed. Overlap: buckle-trap CADs: no geometry overlap with the 132 v1.8 CADs (v1.9 overlap report); the other cases are distinct v1.8 primary tasks; identical CAD files within this set: 0. Astra generation cost for the included attempts: $9.563. Non-CAD+FEM Astra failures (CAD editing, pipe network, frame analysis) are in astra-failure-cases-v0.1.

This public copy

3 cases (A-CYC-Disk-PETGCF, A-CYC-Ring4x6-PETGCF, A-FCVM-PL-Wyk_Rohrstoss_ring) come from sources without a public redistribution licence: their inputs and reference values are withheld (tasks/<id>/WITHHELD.md); the case record and Astra's attempt are kept. Home-directory paths in logs are replaced by <SVG> and ~. Licences per case: LICENSE.md. The full benchmark: https://huggingface.co/datasets/Conditiondark/cadfem-bench.